AUTOMATED DIGITAL IMAGE ANALYSES
FOR ESTIMATING PERCENT GROUND COVER
OF WINTER WHEAT BASED ON OBJECT
FEATURES
Chunjiang Zhao 1,* , Cunjun Li 1 , Qian Wang 1 , Qingyan Meng 2 , Jihua
Wang 1
1
National Engineering Research Center for Information Technology in Agriculture, Beijing,
P. R. China 100097
2
Institute of Remote sensing Applications, Chinese Academy of Sciences, Beijing, P. R. China
*
Corresponding author, Address: National Engineering Research Center for Information
Technology in Agriculture, Beijing 100031, P. R. China, Tel: +86-10-51503411, Email:
[email protected]
Abstract:
Percent ground cover of vegetation is an important parameter for crop
management. An innovational method based on features of objects was
presented to automatically estimate percent ground cover of winter wheat from
digital image analyses. Based on the features of wheat and its background
components, an algorithm was designed to extract the percent ground cover,
and the corresponding program was developed. This method was simple,
labor-and time–saving with high classification accuracy about 90%, and the
method combined the advantages of ISODATA method and maximum
likelihood method. Finally the error source of automatic classification and
scope of application were analyzed, and some approaches of improving
accuracy of classification were discussed.
Keywords:
percent ground cover, image analyses, features extraction, automation
classification
254
1.
Chunjiang Zhao , Cunjun Li , Qian Wang , Qingyan Meng , Jihua
Wang
INTRODUCTION
Percent ground cover of vegetation is an important parameter which both
agronomists and ecologists concern about. Not only does it reflect plant’s
dynamic growth in a long time, but also it is associated with abstraction of
photosynthesis available radiation (APAR) of plant. It can reflect
photosynthesis and transpiration (J E Adans et al., 1977). With the
development of precision agriculture, remote sensing has provided an
important tool for retrieving crop’s water or fertilizer stress, monitoring crop
conditions and offering decision information for the optimized agricultural
production management. Percent ground cover of vegetation is closely
related to remote sensing spectrum index. Stanhill (1972) pointed out that
the differences of spectrum abstraction were due to the amounts of biomass
and percent ground cover of vegetation (Stanhill G et al., 1972). Wanjura
and Hatfield (1987) considered that the remote sensing crop index was
affected by percent ground cover of vegetation more than by dry fresh
biomass or other parameters such as LAI (Wanjura D F et al., 1987). So,
estimating percent ground cover of crop is of great importance in agriculture
cultivation.
The traditional ways (J E Adans et al., 1977; Armbrust D V; Robert P, 1999)
to estimate percent ground cover of crop are as follows:
Agronomists
estimate percent ground cover of vegetation by their experience in field. The
main default of the method is the results are too subjective, and different
persons may have different judging and values, even greatly. A ruler or
electronic device can be used to estimate the values based on the theory of
light capture around noon. The main default of the method is that it is
affected by the weather and the direction when measuring, and it is a laborconsuming method. Percent ground cover of vegetation can be estimated
by observing the pictures taken vertically and by counting the green
elements in the photos.
So the ideal method to estimate percent ground cover of vegetation mainly
includes following procedures: Cheap and easily manipulated equipment
should be adopted. The in-situ data should be accurately and objectively
treated. The method should save time mostly and be restricted the least
when measuring in field. The process should be scarcely disrupted by the
operator (Q Zhou, 2001).
The method to take pictures of crop canopy by digital camera, divide the
image to crop and non-crop (soil and residuals) two classes, and calculate
percent ground cover of vegetation in two-value image arithmetic could be a
good choice among other methods. Along with the expanding market of
digital camera, it is widely used because of the stronger space resolution,
Automated Digital Image Analyses for Estimating Percent Ground
Cover of Winter Wheat Based on Object Features
255
larger storage capacity and cheaper price. Percent ground cover of wheat has
been estimated through alternatively changing contrast and color balance of
digital images by Lukina (Lukina E V et al., 2001) using software
“Micrografx picture publisher”. Percent ground cover of corn has been
estimated through alternatively changing the threshold of red, green, blue
color, the hue, lightness, saturation or principle component analysis of red,
green and blue color by Ewing (Robert P, 1999) using the software “RGBcal
DyEye and RootEdge”. Percent ground cover of pasture has been estimated
through alternatively changing the threshold of tone and saturation by
Richardson using software “SigamaScan pro”. These research results
demonstrate that percent ground cover of crop can be estimated by digital
image analysis technology. However the estimation accuracy is affected by
sunlight condition, the values are more precise in darkness than in sunlight.
Moreover, these methods are labor-and-time consuming with strong
subjectivity and frequent interpretation of the operators, so they are not
satisfying now.
Winter wheat was chosen as materials in the research. With digital camera
to take pictures, the research had adopted relevant analysis technology and
put forward an automated extraction method to estimate percent ground
cover of crop with more accuracy and objectivity. The aim is to offer a
credible and rapid method to estimate percent ground cover of crop.
2.
2.1
EXPERIMENTAL DESIGN AND
METHODOLOGY
Experimental design
When measuring agricultural parameters, the digital image of winter
wheat was taken at the Country Precision Agriculture Demonstration Base in
the growing season in 2002. The experiment site is located in Changping
district ( 40°10 6 to 40°11 2 N, 116°26 3 to 116°27 0 E). The varieties
adopted in the experiment were horizontal “Zhongyou 9507”, erect
“Jingdong 8” and moderate “Jin 9428”.
The factors included in the experiment compromised 4 fertilizer
applications and 4 water strategies in order to study stress variability. There
were 48 plots with 32.4m×30m for each. Four treatments received the
following amounts of N at: 0,150, 300 and 450Kg/hm2. Four water strategies
were applied at rates of 0, 225, 450 and 675 m3/hm2.
256
Chunjiang Zhao , Cunjun Li , Qian Wang , Qingyan Meng , Jihua
Wang
2.2
Field data acquirements
2.2.1
Digital pictures acquirements
The winter wheat in the 48 plots was taken pictures of by Sony camera,
four times before full crop and twice after that. The first four days were
Mar.25th, Apr. 2nd, Apr. 10th and Apr. 18 respectively, leaving the other
two days in May. When taking pictures, the digital camera should be
positioned erectly 1.5m above the ground between 11 o’clock am 12 o’clock.
Because of the weather, some pictures were taken under the shadow of cloud.
The pixel size of the photos is 2228×1712, which were stored in the
computer in the form of JPEG (joint photographic experts group) and the
total number of the image was 192.
As image of the digital camera was derived from the theory of center
projection, which contributed to the distorted image when the field view
angle grows larger, the image fringe has the most distortion. So only the
1284×1284 pixels in the center of the photos were taken to analyze and
compute the percent ground cover of wheat.
2.2.2
LAI index measurements
After the digital pictures were taken of the winter wheat, the
corresponding wheat should be collected thereafter. The dry biomass method
was adopted to measure LAI in Laboratory. 50 to 100 standard leaves were
chosen from each sample to measure leaf area (marked S1) and weigh the
dry weight of these leaves (W1) and the rest leaves for each sample (W2).
And the LAI of each sample can be calculated by the following equation:
LAI=S1× (W1+W2)/W1. The measured values of LAI were revised by CI203 Laser Area Meter finally.
2.3
Methodology
2.3.1
Background-red, green, blue (RGB) and hue, lightness,
saturation (HLS) color space
Digital camera has coupling filters of three different colors, each of the
filter corresponds to a special sensitive spectrum band. The three different
sensitive band are red(R), green (G) and blue (B) in Descartes color space.
Automated Digital Image Analyses for Estimating Percent Ground
Cover of Winter Wheat Based on Object Features
257
Any color in the nature can be represented a point in the space, ranging from
0 to 255.
Another color space is Munsell system, which depicts color with hue (H),
saturation (S) and lightness (L) in accordance to our vision. The space is
defined as a two cone body with two vertexes, one is white and the other is
black. Hue can be represented by ways of surrounding the vertical axis
which itself represents lightness, and the emanative radius along the
horizontal direction toward outside represents saturation (Robert P, 1999;
Kenneth R, 1996; MEI An-xin, 2001). The colors depicted in either space
can be transferred into the other.
2.3.2
Wheat image features
From re-green stage (Mar. 25th) to late jointing stage (Apr. 18th), it is a
fast-growing period for wheat and the percent ground cover increases rapidly.
Before the cover reaches 100%, the height of wheat is low and leaves are
all green without wilted ones. The hot-spot effect isn’t apparent in sunlight.
On the other hand, there are lots of bare land and crop residues. In digital
image, the soil background and crop residues change greatly and the
lightness of certain parts is higher than wheat. From the statistic values of
RGB, variances of the two classes of objects are small. After the cover
reaches 100%, winter wheat grows high and vertical layer has been shaped.
On different layers of wheat leaves, the sunlight conditions vary. Under
strong sunlight, the upper layers have strong hot-spot but the lower layers are
in the shadow of upper layers. Bare land can hardly be seen in this period.
Based on the theory that variances among different objects should be the
least, the image can be classified as five classes as follows:
wheat under
sunlight without hot-spot; wheat under sunlight with hot-spot; wheat in
the shadow; soil under the sunlight; soil in the shadow. There is wax on
the surface of some wheat. Though only wheat and background need to be
considered, however they can’t be simply classified as two classes according
to a certain program because there are lots of transitions between the two
classes in changing sunlight.
2.3.3
Classification theory
The green vegetables appear green because of the strong abstraction in the
red and blue spectrum band region and reflection of most of solar radiation
in the green spectrum region. So the reflected radiation in the green spectrum
is more than that in red and blue spectrum region.
258
Chunjiang Zhao , Cunjun Li , Qian Wang , Qingyan Meng , Jihua
Wang
As to the image before full crop, wheat satisfies the form G R and G B.
However, background (soil, crop residues, etc) covers a large portion in the
image, and some crop residues appear brown or grey brown which also
satisfy the form G R and G B. The residues characters are that the RGB
values have slight differences and the G value is the highest among the
values. So it is necessary to eliminate residue pixels. It is a good way to
distinguish green and brown objects with saturation when the RGB color
space is transferred into HLS color space. In the HLS color space, saturation
60 degrees is kelly and 240 degrees is bottle green. The saturation value of
crop residues is 0-20 degrees and 340-360 degrees, and that of wheat is 60240 degrees.
As far as the image after full crop, wheat cover is above 80%, mainly
around 90%. So wheat in the period can be classified as three sub-classes. (1)
Wheat with hot-spot, it is strongly shined in the daytime and many lightness
values are high even above 0.9. Part of soil under sunlight is relatively
lighter, but the lightness value is less than that of wheat with hot-spot as the
soil is lower than wheat. As a result, the threshold of L values is x L 1, in
which x stands for the highest value of soil or a value above the highest one,
in the paper it is set at 0.9. (2) Wheat under sunlight, the color of it turns
from olivien to bright green, and the corresponding saturation values range
from 90 to 240 degrees. It can’t be misclassificated as soil. (3)wheat in the
shadow, the color of it is green and it satisfies the form R G and R B. In
order to avoid the effects of soil, 60 H 90 is added as the limitation factor.
Maybe there are overlapped parts between the first and the second part, but
the soil has been separated successfully. Finally the wheat image can be got
by combining the three subclasses: (1), (2) and (3) together.
2.3.4
Automated extraction algorithm on percent cover of wheat
According the automated extraction theory, the arithmetic is as follows:
(1)automated extraction arithmetic image before full crop
If the pixels satisfy the following form : R G and R B, 60 H 240 ,
they can be classified as wheat, or else as soil (R, G, B represents red , green,
blue value respectively, and H represents saturation of pixel).
(2)automated extraction algorithm image after full crop
A. Transfer RGB color space into HLS color space, H ranges from 0 to
360 (0 represents red , 120 represents green, 240 represents blue), L and S
range from 0 to 1 (2002).
B. If the pixels satisfy the form 90 H 240, they can be viewed as wheat
and belong to class one as mentioned above correspondingly. The color turns
from olivine to bright blue and covers the corresponding part.
Automated Digital Image Analyses for Estimating Percent Ground
Cover of Winter Wheat Based on Object Features
259
C. As to the pixels left, if they satisfy the form x L 1, they can be viewed
as wheat and belong to class 2 mentioned above correspondingly (x is the
highest value of soil or the value over the highest one, and it is usually set at
0.9).
D. For the rest pixels that are excluded from the above two classes, if the
pixels satisfy the form: R
R G and R B, 60 H 90, they can be viewed as wheat and belong to
class 3 as mentioned above.
2.3.5
Automated extraction program on percent cover of wheat
Based on the algorithm mentioned above, automated extraction program
on percent cover of wheat is developed in IDL Language (MEI An-xin,
2001), which compromises three steps as follows. The first step is to read all
the digital images and files in the same directory by calling READ_JPEG in
the batch processing. The second step is to transfer RGB values to HLS
values by calling COLOR_CONVERT. Using the extraction algorithm of
wheat, digital image before and after full crop can be differentiated pixel by
pixel and classification results (value one represents wheat and value zero
represents background) are stored into array. The third step is to call
IMAGE_STATISTICS to statistically classify the frequencies of one and
zero in the array. Then figure out percent cover of wheat and export the
value into files.
The program uses figure-user interface which can be easily manipulated,
and the main procedures include: (1) put the image that needs to be treated
with in the same directory; (2) run the program and input directory name; (3)
the program will be automatically searching for all the images in the
directory and classifies them individually; (4) export the results, each image
will engendering a chart that contains classification results information and
the percent cover of wheat data in all the images will be exported into a file.
3.
3.1
EXPERIMENT RESULTS AND EVALUATION
Automated extraction accuracy on percent cover of
wheat
Computer digital image classification includes supervised classification
and unsupervised classification. Supervised classification firstly needs to
choose training sample from manual work and the accuracy is closely
260
Chunjiang Zhao , Cunjun Li , Qian Wang , Qingyan Meng , Jihua
Wang
correlated with the amounts and the degree of representation of the samples
chosen. The unsupervised classification needn’t choose training sample
because it combines and classifies the pixels by their similarity, so it is an
automated method.
The most popular methods in the research are unsupervised
classification—ISODATA method, supervised classification-maximum
likelihood method (ML), and the automated extraction method suggested in
the paper. And six images were randomly chosen in the experiment to
classify wheat and extract the percent cover of wheat, the accuracy of the
three methods were evaluated by the observing the digital pictures visually.
Among the pictures, the former four were taken of before full crop and the
two latter pictures are after full crop, which are shown in Figure 1.
Fig.1:Six cases of three classification results and their comparisons
ISODATA method and maximum likelihood method were run with
software packet ENVI3.5 (2002). Before applying the maximum likelihood
method in the whole image, the evenly distributed training samples of wheat
and soil background should be chosen manually and respectively.
The classification results of the six images applying the three
classifications mentioned above are shown in Figure 1.
One hundred pixels randomly chosen in the six images were judged
visually and classified as wheat and soil then compared the results of the
three methods with the visual judgment. Rusesll’s error matrix (Russell G
Congalton, 1991) provides reference to assess the accuracy of the three
methods in the paper. The error matrix for accuracy assessment as Table 1.
Automated Digital Image Analyses for Estimating Percent Ground
Cover of Winter Wheat Based on Object Features
261
Table 1 The error matrix for classification accuracy assessment
Reference data
(from eye judgment)
wheat
soil
total
Classification data
wheat
a
b
number of wheat
soil
c
d
number of soil
Note: a: the number of wheat pixels that have been classified as wheat in
the reference data; b: the number of wheat pixels that have been classified as
oil in the reference data; c: the number of soil pixels that have been
classified as wheat in the reference data; d: the number of soil pixels that
have been classified as soil in the reference data. The overall classification
accuracy can be calculated by the form: (a +d)/ (a + b+ c+ d) ×100%.
The classification accuracy of the six images with three different methods
is shown in Table 2, respectively.
Table 2. Overall accuracy(%) of six image with three classification methods
Isodata
the maximum likelihood
the automated extraction method
image 1
63
90
87
image 2
55
91
95
image 3
54
89
88
image 4
48
91
85
image 5
68
85
93
image 6
45
92
93
average
55.5
89.7
90.2
Table 2 indicates that ISODATA method has the lowest classification
accuracy (45%-68%, 55.5% on average), and the maximum likelihood
method and the automated extraction method both have high accuracy with
little difference.
The comparisons of the three methods are shown in Table 3. The theories
of them are different. In terms of choosing training samples, objectivity,
batch processing, work amounts, processing time and individual
participation of the three classifications, automated extraction method and
ISODATA method are very similar, simple and fast. However the maximum
likelihood method is labor-and-time consuming. With respect to
classification accuracy, the maximum likelihood method and automated
extraction method have higher classification accuracy than ISODATA
method. So the automated extraction method based on the digital image
features combines the advantage of ISODATA method and maximum
likelihood method, it can provide the estimation rapidly with the least work
amounts and high accuracy.
262
Chunjiang Zhao , Cunjun Li , Qian Wang , Qingyan Meng , Jihua
Wang
Table 3. The compare of advantage and disadvantage with three classification methods
ISODATA method
maximum likelihood method
automated extraction method
theory
the overall statistic based on
a whole image RGB, the
iterative unsupervised
classification by its average
value and dispersion
the RGB statistic based on the
training sample, classify the
parts out of training samples in
the image according to
maximum likelihood principle
classification based on the
characters of wheat and soil,
their features in the RGB and
HLS color space, and the
sunshine effect on imaging
Choosing
training
sample
unnecessary
absolutely necessary, every
image has its corresponding
training sampl
unnecessary
objectivity
strong objectivity
strong subjectivity, correlated
with training sample
strong objectivity
Batch
processing
enable
unable
enable
Work
amounts
small
heavy
small
Processing
time
short
long
short
Individual
participation
occasionally
Frequently
occasionally
accuracy
very low
high
high
General
evaluation
bad
fine
fine
3.2
Evaluation of the error and applicability of the
automated extraction method
As to the digital photos, a small part of wheat may be easily
misclassificated as soil because of the strong similarity between soil and
wheat in color and brightness, especially most of the pictures were taken in
sunlight. In addition, the transition from wheat to background can easily be
confused due to the mixed pixels in a digital picture.
Were there absolutely not hot-spot effect on wheat leaves (under obscure
sunlight or cloudy whether), the classification accuracy would increase
somewhat.
As to the digital picture after full crop, the bottom wheat was covered by
the top wheat and received less sunshine. Probably some of the leaves can be
misclassificated as soil background due to the dark environment.
The accuracy of the latter two images after full crop was higher than that
of the former four images before the full crop. Soil received less sunshine
after full crop and it can be distinguished from wheat easily because they
were very different from each other by the color and lightness.
The soil type in the research was alluvial soil. With different soil moisture
and under different sunlight the soil appeared grey, French grey, filemot,
buff and khaki in the digital image. The research results can be applied to
wheat grown on soil types of black, grey, khaki etc. series. Further
Automated Digital Image Analyses for Estimating Percent Ground
Cover of Winter Wheat Based on Object Features
263
discussion and validation of the applicability of the red or other soil types
were still necessary.
By adjusting some parameters, the automated extraction method can be
applied to extract percent cover of soybean, clover, cotton, corn and rice and
to monitor crop dynamic growth.
3.3
Estimation of leaf area through percent cover of
wheat
When regressed against percent cover of wheat before full crop extracted
by using digital image and the measured LAI , it is found that the two
showed exponential relationship, which reached highly significant level, and
the multiple correlation coefficient was 0.743. It is consistent with Toby’s
research result (Toby N, 1997).
4.
DISCUSSION
This paper brings forward the automated extraction method of wheat
percent cover using digital image analysis technology to process the pictures
taken with digital camera. The method eliminates operator’s subjectivity and
it has high classification accuracy and satisfying results. In addition, as the
estimation accuracy is affected less by the sunshine conditions, the satisfying
results retain whether in sunshine or in shadow.
Furthermore, the LAI before full crop could be estimated by percent cover
of wheat from the digital image.
As to the wheat image before full crop, part of it may be misclassificated
as soil due to strong sunshine and hot-spot phenomenon. Part of the pixels of
the transition between wheat and background may be misclassificated
because of mixed pixels. The bottom leaves received less sunshine as the top
leaves, may be misclassified after full crop.
The new algorithm in the paper is based on single pixel in wheat digital
image without considering its relationship with surrounding pixels and prior
knowledge, such as that wheat was planted along rows, wheat is composed
of many pixels while the isolated pixel indefinitely isn’t wheat, and provided
there was hallow hole on the physical surface of wheat leaves (generally the
hot-spot part), the hallow hole would be wheat. In conclusion, in order to
improve estimation accuracy, physical morphologic algorithm should be
combined with and prior knowledge should be added for further research.
264
Chunjiang Zhao , Cunjun Li , Qian Wang , Qingyan Meng , Jihua
Wang
ACKNOWLEDGEMENTS
This research was supported by National Natural Science Foundation of
China (40701120), National High Tech R&D Program of China
(2006AA12Z138, 2007AA10Z201), and Excellent Talent Training
program(20071D0200500046).
REFERENCES
Armbrust D V. Rapid measurement of crop canopy cover[J].Agron.J, 82:1170-1171
J E Adans, G F Arkin. A light interception method for measurement row copy ground
cover[J].Soil Science of America Journal, 1977, 789-791
Kenneth R. Castleman.Digital Image Processing[M], Prentice Hall, Inc, 1996
Kuusk A. The hot-spot effect in plant canopy reflectance, Photon-Vegetation
interactions[J].Application in optical remote sensing and Plant Ecology, Springer-Verlag,
New York, USA, 1991,139-159
Lukina E V, M L Stone, W R Raun. Estimating vegetation coverage in wheat using digital
images[J].J. Plant Nutr, 1999,22:341-350
M D Richardson, D E Karcher, C Purcell. Quantifying turfgrass cover using digital image
analysis[J].Crop sci, 2001, 41:1884-1888
MEI An-xin.Introduction to remote sensing [M]. Beijing: Chinese high education
press,2001(in Chinese)
Q Zhou. Automated rangeland vegetation cover and density estimation using ground digital
images and a spectral-contextual classifer[J].Int.J.Remote Sensing, 2001, 22(17):3457-3470
Robert P. Ewing and Robert Horton, Quantitative color image analysis of agronomic
images[J].Agron.J, 1999,91:148-153
Russell G Congalton. A Review of Assessing the Accuracy of Classifications of Remotely
Sensed Data[J]. REMOTE SENS,ENVIRON,1991 37:35-46
Stanhill G, U Kafkafi, M Fuchs,et al. The effect of fertilizer application on solar reflectance
from a wheat crop[J].Israel J.agric.Res,1972, 22(2):109-118
Toby N. Carlson, On the Relation between NDVI, Fractional Vegetation Cover, and Leaf
Area Index[J]. Remote sensing of Environment, 1997, 2:241-25
Wanjura D F, J L Hatfield. Sensitivity of spectral vegetative indices to crop
biomass[J].Transactions of the ASAE, 1987, 30(3):811-816